Initializing of models from checkpoints/pretrained models has gotten a bit crazy

[dlwh]

We have a lot of ways of initializing models right now:

1. from scratch
2. load the whole state from a levanter checkpoint of the current run (default if present)
3. loading the whole state from a levanter checkpoint of a different run (trainer.load_checkpoint_path)
4. loading just the model weights from a levanter checkpoint (trainer.initialize_from)
5. loading the weights and optimizer state, with an eye towards changing the data (.initialize_from)
6. ladoing from hf --initialize_from_hf

This is a bit insane and convoluted. We should come up with some rational strategy here.